In the Figure below (on synthetic data), you can see how the model learns to align predictions with the target quantization over training steps.

To ensure the predictions are meaningfully different (not just slight variations), we use a Winner-Takes-All training strategy that updates the best-performing prediction per example. This leads to quantization properties, where the predictions serve as representative prototypes of the future

In our paper we introduce TimeMCL, a method designed to predict multiple plausible futures for time series data.

TimeMCL builds on a technique called Multiple Choice Learning, which trains a model to generate a diverse set of predictions rather than a single outcome.

When we try to predict what might happen in the future based on past data, we often find that there isn’t just one “right” answer — there could be several possible future scenarios.

Interested in time series forecasting or data uncertainty quantification?
Check out our latest paper with Adrien Cortés at @icmlconf !

Paper: arxiv.org/abs/2506.05515
Code: github.com/Victorletzel...
Poster #2211 , Tue 15 Jul 11 a.m. PDT East
#timeseries #quantization #uncertainty #icml2025

🚗 Ever wondered if an AI model could learn to drive just by watching YouTube? 🎥👀

We trained a 1.2B parameter model on 1,800+ hours of raw driving videos.

No labels. No maps. Just pure observation.

And it works! 🤯

🧵👇 [1/10]

Inferring 3D human poses from video is highly ill-posed because of depth ambiguity.

Our work accepted to #NeurIPS2024, ManiPose, gets one step closer to solving this, by leveraging prior knowledge about poses topology and cool multiple-choice learning techniques.